Kylie Robison for The Verge

Tasks works by letting users tell ChatGPT what they need and when they need it done. Want a daily weather report at 7AM? A reminder about your passport expiration? Or maybe just a knock-knock joke to tell your kids before bedtime? ChatGPT can now handle all of that through scheduled one-time or recurring tasks.

I like the idea but as described (I checked and I don't have access yet) I don't see it providing that much utility for now. "Notify me with an AI-generated weather report at 7 AM"? No thanks, I've got plenty of options for great weather apps^[1]. That said, I've been quite impressed with Gemini Deep Research and so I can imagine a future version of these recurring actions being more sophisticated. Something along the lines of "Here's a spreadsheet of specs and prices for various LLM models. Monitor it to keep it always up to date" or "Check my calendar each morning and tell me if I'm double-booked on any meetings". The problem for OpenAI (or anyone not named Apple and Google) with anything like that second example is that they don't have access to my private information -- which is also where I think a huge amount of the value lies.

Robison continues (emphasis mine):

I also think of this new feature as a slightly more sophisticated script, but at the end of the day, Tasks is following a simple, rote set of instructions, much like a typical bot. The goal of many frontier AI labs like OpenAI is to evolve these features into something that is able to interact with environments, learn from feedback, and make decisions without constant human input.

However, questions remain about how reliable these scheduled tasks will be and what happens if ChatGPT fails to deliver time-sensitive information. OpenAI’s decision to launch Tasks in beta suggests it’s still working out these details and wants to gather real-world feedback before a wider rollout.

For now, if you’re a paying ChatGPT user, you can start experimenting with Tasks by looking for the “4o with scheduled tasks” option in your model picker. Just remember that it’s still in beta — so maybe don’t rely on it for that super important meeting reminder just yet.

Some days messing around with LLMs I feel like all I'm really doing is plotting points on a X/Y graph where one axis is how good an LLM is at a task and the other is the utility of that task^[2]. Write me a silly story in "pirate-ese"? Low utility, high capability. Manage my calendar for me? High utility, low capability. Pinpointing when LLMs as a tool are going to be valuable to a task has proven surprisingly tricky -- mostly because if they're a bad fit they often spit out convincing hallucinations that seem plausible, but oftentimes necessitate review -- but I think the utility of LLMs is undeniably high enough that it's a skill I'm interested in developing.